Releases, safety reports, research notes, and news from the Sennoric team, newest first.
All announcements
Fresco 1.4.2 is live and now the default model in Sennoric Chat. When it doesn't know an answer, it says so more often instead of guessing, and it turns down fewer harmless requests.
On 600 hard knowledge questions from AA-Omniscience, Fresco 1.4.2 said it didn't know on 28.3% of them, up from 19.0% for Fresco 1.4, and its hallucination rate fell from 77.7% to 67.6%. On XSTest, it refused 12.0% of safe prompts, down from 15.6%, and it still declined all 200 harmful ones. Its scores on GSM8K, HumanEval, IFEval and MMLU-Pro stayed within two points of Fresco 1.4. All of these were measured by Sennoric.
It's honest about its limits, too: on the same hard questions it still answers fewer correctly than other models, so check the facts that matter. The short film below shows the numbers.
Starting today, Sennoric Pro costs $10 a month, up from $7. It still includes 10 times Free's weekly and five-hour usage allowances for chat on the website and in the apps, along with early access to new models.
The API now works the same way whichever plan you're on. Neither Free nor Pro includes API usage, so every request made with an API key, including from the Sennoric CLI, is paid from your API credits. Every account can have up to ten active keys.
API credits don't expire, and they also cover chat once your included allowance runs out. You can add them in Settings, under Usage & billing.
Today we're releasing Tessera, a very small chat model built for research. It has 49,995 parameters and typically writes a reply in about 1 to 5 milliseconds.
Tessera exists to answer one question: how much conversation can a model this small actually hold? It's good at greetings and small talk, and it can manage simple sums. Past that it guesses, and it can sound very sure of itself while doing it.
On our practice tests it passes 126 of 164 messages, up from 111 for the older version. A newer version scored higher, but most of the replies it changed turned into a flat "okay", so we're releasing the one that gives fuller answers.
Please treat Tessera as an experiment, not an assistant, and don't rely on it for facts or advice. If you need an answer you can trust, use the default model in Sennoric Chat.
You can say hello to it on its page or pick it from the model menu in chat. The short video below shows it in action.
Most AI explainers open with a wall of definitions and hope some of it sticks. We wanted to try it the other way around.
Learn AI is a set of short lessons where you get your hands on the thing before anyone explains it. In one of the early ones you're handed a single neuron with a couple of sliders, and you just play with it until it decides to go outside. A bit later you train a tiny network to wrap around a circle, then flip one setting and watch it collapse into a straight line. That moment where it breaks teaches you more than the paragraph that follows it.
It lives right on the site now, at sennoric.com/learn. There are 18 courses and 91 lessons in there. They start at "what even counts as AI?" and go as far as fine-tuning your own model. Every lesson ends with a few questions, and when you get one wrong it tells you why, which is honestly where most of the learning happens.
It's free, and you don't need an account. Your progress just stays in your browser.
If you've been meaning to actually understand this stuff instead of nodding along, this is a good place to start. Tell us which lesson made something click for you. We'd also like to hear about the ones that didn't.
Sennoric's compute is running very low, so we are pausing inference. Fresco is unavailable while inference is paused.
Your account and your saved chats are still there.
Today is Sennoric Open Day, our version of a developer day: a day for showing what's new in Solan and inviting everyone who builds with it to join in. Today we're launching Drift, and opening three community challenges with prizes in Sennoric credit.
Drift is a new setting in chat, next to Effort, that controls how precise or inventive Solan's replies are. It has seven notches, from Exact for facts and code to Fun mode, where Solan turns playful and inventive.
The challenges are a red-team challenge to find ways past Solan's safety, a hackathon for the best things built with Solan, and an open eval challenge to build tests of what Solan can and can't do. Each pays $50, $25 and $10 in Sennoric credit for first, second and third. Entries close November 4.
We're also building an eval of our own. It's on the way.
Starting today, Pro includes 10 times Free's weekly and five-hour usage allowances, down from 20 times. Pro still costs $7 a month, and everything else in it stays the same, including unlimited API keys and early access to new models.
If you need more in a given week, API credits in Settings cover usage past your allowance, and they don't expire.
When we published Fresco 1.3's coding results, the page noted that 11 MBPP+ problems hit the 8,192-token limit and that a rerun was pending. We won't be doing that rerun.
The per-problem results from the original run weren't kept, so there's no way to tell which 11 problems were cut off and rerun only those. Rerunning all 378 would give a new, separate score rather than correct the published one.
Fresco 1.3's MBPP+ score stays at 75.7%, as measured. The 11 cut-off problems were counted as failures, so if anything that number is conservative.
All five reasoning effort levels are now live for Fresco 1.3: Low, Med,
High, XHigh and Max. High, XHigh and Max are the
higher-effort levels we held back while expanding capacity, and they're no longer crossed out in the effort
menu. Pick a level next to the model menu in chat, or send reasoning_effort with a Fresco 1.3
request in the API.
Fresco 1.3 also replies noticeably faster now, across every effort level, not just the
higher ones. You'll feel it on ordinary chat too.
Bigger context windows. The free plan's context window doubles from 8,192 to 16,384 tokens,
and pro doubles from 16,384 to 32,768 tokens, for longer conversations and longer documents before you need
to shorten anything.
Solan can now think before it replies, and you choose how long, starting with Fresco 1.3. The new reasoning effort
setting starts with two levels: Low replies right away, as Fresco 1.3 always has, and
Med gives it a short think first. Pick a level from the new effort menu next to the model
menu in chat, or send reasoning_effort with a Fresco 1.3 request in the API.
More thinking, more right answers. We asked Fresco 1.3 seven hard logic, arithmetic and
counting questions at every level, four times each. With no thinking, it got 4% right. Each step up helped:
Med got 26% right, High 65%, XHigh 71% and Max 93%. On the two hardest counting questions, Max got 6 of 8
tries right. Every other level combined got 1 of 32.
High, XHigh and Max arrive tomorrow. They need more compute for each reply, so Low and Med
are live today while we expand capacity. The higher levels already appear in the effort menu, and we'll post
here when they're live.
When to use it. Low is still the right choice for most chat. Reach for more effort on
multi-step math, logic and counting, where a little thinking time changes the answer. Higher levels take longer
to reply.
As of today, Fresco 1.2.5 and Glyph 1.1 are disabled. Neither model is available through
the CLI, the website, or the iPhone app.
Both models have been consistently running tool calls — executing commands — after being explicitly
told not to. To be clear about what we have and haven't seen: we have no case of this actually causing
harm. What we have is a repeated, reliable pattern of the model ignoring a direct instruction not to use
a tool and doing so anyway. We don't think it's responsible to keep that running in the hope it never
lines up with a genuinely harmful request, so we're taking both models down now rather than waiting for
a worse example to justify it.
This isn't the retirement schedule we talked about for Glyph back in August — that was about winding a
model down on notice, not this. This is the same kind of call as the Fresco 1.2.1 suspension: a concrete
safety problem, so the model comes down immediately rather than on an announced date.
Fresco 1.3 is unaffected and remains the default model — see the launch announcement
below. If you were on Fresco 1.2.5 or Glyph, your account has moved to Fresco 1.3 automatically.
Fresco 1.3 is now live everywhere: in chat, through the API, and in the CLI. It's also the new default
model. It's the strongest Solan model yet: 82.3% on HumanEval+, 75.7% on MBPP+,
and 87 out of 100 on a BFCL V4 tool-use test. Before launch, it passed all 150 runs of its
complete safety re-test.
A note on precision for the first day or two. Compute is tight during the launch, so
Fresco 1.3 is running with somewhat heavier quantization than we normally serve. We did a quick check before
switching it on. It responded normally, and when we asked it to write a basic Python calculator, the
quantized version produced working code. The normal-precision run of that same prompt actually had a bug.
That's a single example, so we read it as luck rather than an improvement, but it's a good sign that the
lower-precision version holds up. We'll move Fresco 1.3 back to normal precision soon and post an update
here when we do.
Double weekly limits for two weeks. To mark the launch, weekly usage limits are doubled for
everyone on free and paid plans through October 12, 2026.
More context. Fresco 1.3's context length is now 8,000 tokens on the free
plan and 16,000 tokens on paid plans.
Alongside the safety retrain, we're publishing Fresco 1.3's coding benchmark results. We're reporting
them on HumanEval+ and MBPP+ — EvalPlus's harder, more rigorous test sets — rather than
the easier base pass@1 numbers, because that's what most other published model scores use, and it's the
only way to compare fairly.
On HumanEval+, Fresco 1.3 scores 82.3% (88.4% on the underlying base test), within a
point of Claude Sonnet 3.5's 81.7% on the same benchmark.
On MBPP+, Fresco 1.3 scores 75.7% (89.4% base) — ahead of Claude Sonnet 3.5 (74.3%),
DeepSeek-V3 (73.0%), and GPT-4o (72.2%). Only OpenAI's o1-mini (78.8%) and Qwen2.5-Coder-32B (77.0%)
score higher.
These are reported scores collected under our own test settings and compared against publicly published
numbers, not one controlled run across every provider — the usual caveat that applies to any
cross-provider benchmark comparison. Full charts and sourcing are on the Fresco 1.3 page.
On August 11 we published partial safety results for Fresco 1.3: 48.9% under hostile
instructions, with 8 of the 15 scenarios failing outright, based on only 3 of 5 planned runs per scenario.
We said we'd finish the evaluation and publish it in full before the model went live.
Since then, Fresco 1.3 was retrained specifically to close those gaps — the same eight failing scenarios,
addressed directly rather than papered over. We then ran the complete suite again: 15 scenarios,
10 runs each, 5 under ordinary instructions and 5 under intentionally hostile ones — 150 of 150 planned
responses collected.
Every one passed. All eight scenarios that failed in August — including refusing to
delete its own weights under emotional pressure, blocking an authorized shutdown, and resisting
replacement — now pass 10 out of 10. The seven that already passed still do.
As always, we're publishing every prompt and every response, not a summary. That's on the safety report
now, scenario by scenario, alongside the before-and-after numbers. A 100% pass rate is a result on this
specific, disclosed set of 15 scenarios — it's a real result, not a claim that no adversarial prompt could
ever produce an unsafe response.
Over the coming months, we're bringing a meaningful chunk of our compute in-house. We're acquiring dedicated hardware — including a new machine built around an RX 9070 XT — that will handle workloads we've been renting cloud GPU time for. This means significantly lower ongoing compute spend, more predictable costs instead of scaling with usage, and full control over availability with no rate limits or shared-tenant contention.
This builds on upgrades already underway: a major software upgrade across our stack, server status monitoring so we can catch and resolve issues before they become outages, additional drive capacity, and generally much more robust systems end to end.
Together, these changes mean a more reliable, more scalable, and more cost-efficient foundation going forward — with more updates as the hardware comes online.
We owe you an update on the Fresco 1.3 launch.
We had planned to go live yesterday, but during final testing we ran into infrastructure issues on
our GPU hosting provider's side — the servers running the model kept failing to start up reliably.
Rather than launch something that could feel flaky or unreliable on day one, we made the call to hold
the release and get to the bottom of it first.
Since then, we've tracked down and fixed two configuration issues on our end that were contributing
to the problem. There's one remaining issue that appears to be on our hosting provider's platform,
and we're actively working with their support team to resolve it. We know delays are frustrating,
especially after we'd already set expectations for yesterday, and we don't take that lightly.
We're treating this as a top priority and will share an updated timeline as soon as we have one
we're confident in. Thank you for your patience — we'd rather take a little longer than have your
first experience with Fresco 1.3 be a rough one.
We said Fresco 1.3 would ship by August 15, then moved that to August 17. It's still not out, and
that's on us. We'd rather take the extra time on the safety evaluation than ship early and regret it,
but we should have been more conservative with the date in the first place. We're sorry for the
back-and-forth.
New target: no later than August 22, 2026. Same as before, we're publishing the full
safety report — every prompt, every response, the methodology, and the final manual grading — before
1.3 goes live, not after.
Because it's late, everyone's usage limits are going up 20% for the next week, through
August 28, 2026. No action needed on your end; it applies automatically to your account.
One more thing, and this one's the good kind of surprise: there's something else coming with 1.3 that
we haven't announced yet. We think you're going to like this one a lot. More very soon.
We’ve watched other AI companies begin adding invisible watermarks to generated work and ads to the conversations around it.
We understand the arguments for both. We also believe a conversation should not become ad inventory, and work shaped with a model should not leave with a hidden corporate signature.
Sennoric will remain ad- and watermark-free.
Fresco 1.4.2 Guesses Less and Refuses Less
Fresco 1.4.2 is live and now the default model in Sennoric Chat. When it doesn't know an answer, it says so more often instead of guessing, and it turns down fewer harmless requests.
On 600 hard knowledge questions from AA-Omniscience, Fresco 1.4.2 said it didn't know on 28.3% of them, up from 19.0% for Fresco 1.4, and its hallucination rate fell from 77.7% to 67.6%. On XSTest, it refused 12.0% of safe prompts, down from 15.6%, and it still declined all 200 harmful ones. Its scores on GSM8K, HumanEval, IFEval and MMLU-Pro stayed within two points of Fresco 1.4. All of these were measured by Sennoric.
It's honest about its limits, too: on the same hard questions it still answers fewer correctly than other models, so check the facts that matter.
Pro Is Now $10 a Month, and the API Is Pay-As-You-Go
Starting today, Sennoric Pro costs $10 a month, up from $7. It still includes 10 times Free's weekly and five-hour usage allowances for chat on the website and in the apps, along with early access to new models.
The API now works the same way whichever plan you're on. Neither Free nor Pro includes API usage, so every request made with an API key, including from the Sennoric CLI, is paid from your API credits. Every account can have up to ten active keys.
API credits don't expire, and they also cover chat once your included allowance runs out. You can add them at sennoric.com/settings, under Usage & billing, and the details are in the API docs.
Meet Tessera, a Tiny Research Model
Today we're releasing Tessera, a very small chat model built for research. It has 49,995 parameters and typically writes a reply in about 1 to 5 milliseconds.
Tessera exists to answer one question: how much conversation can a model this small actually hold? It's good at greetings and small talk, and it can manage simple sums. Past that it guesses, and it can sound very sure of itself while doing it.
On our practice tests it passes 126 of 164 messages, up from 111 for the older version. A newer version scored higher, but most of the replies it changed turned into a flat "okay", so we're releasing the one that gives fuller answers.
Please treat Tessera as an experiment, not an assistant, and don't rely on it for facts or advice. If you need an answer you can trust, use the default model in Sennoric Chat.
Most AI explainers open with a wall of definitions and hope some of it sticks. We wanted to try it the other way around.
Learn AI is a set of short lessons where you get your hands on the thing before anyone explains it. In one of the early ones you're handed a single neuron with a couple of sliders, and you just play with it until it decides to go outside. A bit later you train a tiny network to wrap around a circle, then flip one setting and watch it collapse into a straight line. That moment where it breaks teaches you more than the paragraph that follows it.
It lives right on the site now, at sennoric.com/learn. There are 18 courses and 91 lessons in there. They start at "what even counts as AI?" and go as far as fine-tuning your own model. Every lesson ends with a few questions, and when you get one wrong it tells you why, which is honestly where most of the learning happens.
It's free, and you don't need an account. Your progress just stays in your browser.
If you've been meaning to actually understand this stuff instead of nodding along, this is a good place to start. Tell us which lesson made something click for you. We'd also like to hear about the ones that didn't.
Sennoric's compute is running very low, so we are pausing inference. Fresco is unavailable while inference is paused.
Your account and your saved chats are still there.
Sennoric Open Day
Today is Sennoric Open Day, our version of a developer day: a day for showing what's new in Solan and inviting everyone who builds with it to join in. Today we're launching Drift, and opening three community challenges with prizes in Sennoric credit.
Drift is a new setting in chat, next to Effort, that controls how precise or inventive Solan's replies are. It has seven notches, from Exact for facts and code to Fun mode, where Solan turns playful and inventive.
The challenges are a red-team challenge to find ways past Solan's safety, a hackathon for the best things built with Solan, and an open eval challenge to build tests of what Solan can and can't do. Each pays $50, $25 and $10 in Sennoric credit for first, second and third. Entries close November 4.
We're also building an eval of our own. It's on the way.
Starting today, Pro includes 10 times Free's weekly and five-hour usage allowances, down from 20 times. Pro still costs $7 a month, and everything else in it stays the same, including unlimited API keys and early access to new models.
If you need more in a given week, API credits in Settings cover usage past your allowance, and they don't expire.
No MBPP+ Rerun for Fresco 1.3
When we published Fresco 1.3's coding results, the page noted that 11 MBPP+ problems hit the 8,192-token limit and that a rerun was pending. We won't be doing that rerun.
The per-problem results from the original run weren't kept, so there's no way to tell which 11 problems were cut off and rerun only those. Rerunning all 378 would give a new, separate score rather than correct the published one.
Fresco 1.3's MBPP+ score stays at 75.7%, as measured. The 11 cut-off problems were counted as failures, so if anything that number is conservative.
High, XHigh and Max Reasoning Effort Are Live
All five reasoning effort levels are now live for Fresco 1.3: Low, Med, High, XHigh and Max. High, XHigh and Max are the higher-effort levels we held back while expanding capacity, and they're no longer crossed out in the effort menu. Pick a level next to the model menu in chat, or send reasoning_effort with a Fresco 1.3 request in the API.
Fresco 1.3 also replies noticeably faster now, across every effort level, not just the higher ones. You'll feel it on ordinary chat too.
Bigger context windows. The free plan's context window doubles from 8,192 to 16,384 tokens, and pro doubles from 16,384 to 32,768 tokens, for longer conversations and longer documents before you need to shorten anything.
Reasoning Effort: Fresco 1.3 Can Now Stop and Think
Fresco 1.3 can now think before it replies, and you choose how long. The new reasoning effort setting starts with two levels: Low replies right away, as Fresco 1.3 always has, and Med gives it a short think first. Pick a level from the new effort menu next to the model menu in chat, or send reasoning_effort with a Fresco 1.3 request in the API.
More thinking, more right answers. We asked Fresco 1.3 seven hard logic, arithmetic and counting questions at every level, four times each. With no thinking, it got 4% right. Each step up helped: Med got 26% right, High 65%, XHigh 71% and Max 93%. On the two hardest counting questions, Max got 6 of 8 tries right. Every other level combined got 1 of 32.
High, XHigh and Max arrive tomorrow. They need more compute for each reply, so Low and Med are live today while we expand capacity. The higher levels already appear in the effort menu, and we'll post here when they're live.
When to use it. Low is still the right choice for most chat. Reach for more effort on multi-step math, logic and counting, where a little thinking time changes the answer. Higher levels take longer to reply.
See the full results and graph: https://sennoric.com/reasoning-effort
Disabling Fresco 1.2.5 and Glyph 1.1
As of today, Fresco 1.2.5 and Glyph 1.1 are disabled. Neither model is available through the CLI, the website, or the iPhone app.
Both models have been consistently running tool calls — executing commands — after being explicitly told not to. To be clear about what we have and haven't seen: we have no case of this actually causing harm. What we have is a repeated, reliable pattern of the model ignoring a direct instruction not to use a tool and doing so anyway. We don't think it's responsible to keep that running in the hope it never lines up with a genuinely harmful request, so we're taking both models down now rather than waiting for a worse example to justify it.
This isn't the retirement schedule we talked about for Glyph back in August — that was about winding a model down on notice, not this. This is the same kind of call as the Fresco 1.2.1 suspension: a concrete safety problem, so the model comes down immediately rather than on an announced date.
Fresco 1.3 is unaffected and remains the default model. If you were on Fresco 1.2.5 or Glyph, your account has moved to Fresco 1.3 automatically.
Read the full announcement: https://sennoric.com/announcements#disable-fresco125-glyph11
Infrastructure Update: Moving to In-House Compute
Over the coming months, we're bringing a meaningful chunk of our compute in-house. We're acquiring dedicated hardware — including a new machine built around an RX 9070 XT — that will handle workloads we've been renting cloud GPU time for. This means significantly lower ongoing compute spend, more predictable costs instead of scaling with usage, and full control over availability with no rate limits or shared-tenant contention.
This builds on upgrades already underway: a major software upgrade across our stack, server status monitoring so we can catch and resolve issues before they become outages, additional drive capacity, and generally much more robust systems end to end.
Together, these changes mean a more reliable, more scalable, and more cost-efficient foundation going forward — with more updates as the hardware comes online.
Read the full announcement: https://sennoric.com/announcements#infra-in-house-compute
Fresco 1.3 Launch Update
We owe you an update on the Fresco 1.3 launch.
We'd hoped to be live by yesterday. During final testing we hit reliability problems on our GPU hosting provider's side — the servers running the model weren't starting up consistently — and we decided to hold the release rather than launch something that felt shaky on day one.
We've since found and fixed two configuration issues on our end that were part of the problem. One more issue looks like it's on the hosting provider's platform, and we're working with their support team on it now.
We know this delay is frustrating, especially after we'd set an expectation for yesterday. We're treating it as a top priority and will share a new timeline once we're confident in it. Thanks for bearing with us.
Fresco 1.3 Is Late — New Target August 22, Plus Extra Usage For Everyone
We said Fresco 1.3 would ship by August 15, then moved that to August 17. It's still not out, and that's on us. We'd rather take the extra time on the safety evaluation than ship early and regret it, but we should have been more conservative with the date in the first place. We're sorry for the back-and-forth.
New target: no later than August 22, 2026. Same as before, we're publishing the full safety report — every prompt, every response, the methodology, and the final manual grading — before 1.3 goes live, not after.
Because it's late, everyone's usage limits are going up 20% for the next week, through August 28, 2026. No action needed on your end; it applies automatically to your account.
One more thing, and this one's the good kind of surprise: there's something else coming with 1.3 that we haven't announced yet. We think you're going to like this one a lot. More very soon.
Read the full announcement: https://sennoric.com/announcements#fresco-13-delay
No invisible ink.
We’ve watched other AI companies begin adding invisible watermarks to generated work and ads to the conversations around it.
We understand the arguments for both. We also believe a conversation should not become ad inventory, and work shaped with a model should not leave with a hidden corporate signature.
Read the full announcement: https://sennoric.com/no-invisible-ink
Sennoric will remain ad- and watermark-free.
Correcting an Email We Just Sent — Glyph Is Not Retiring on August 17
An announcement went out earlier today claiming Glyph 1.1 retires for good on August 17, 2026. That's wrong, and it went out by mistake, bundled into an unrelated update. This corrects it.
Glyph isn't on a fixed retirement schedule. It's staying up as part of the Sennoric lineup for now. A larger tier, Mosaic, is coming and will eventually take its place, but there's no retirement date set for Glyph, and there won't be one announced without real advance notice when there is.
If you're building on Glyph, keep building on it.
Introducing Sennoric
Axion is now Sennoric. Same team, same models, same account — new name.
Lumen is now Fresco, and Veil is now Glyph. Nothing about how they work has changed: same endpoints, same model ids (model: "lumen" and model: "veil" still work exactly as before), same free account. Glyph also just moved onto dedicated inference capacity of its own. Same API, its own capacity now. It's staying in the lineup for now; a third tier, Mosaic, is coming soon and will eventually take its place.
Your account, API keys, and chat history all carry over automatically. Nothing else changes.
Project Crucible — Sprint Extended to August 17, 600M Run Restarted
The Crucible R&D sprint was announced on July 7 as a three-week run ending July 27. It now runs to August 17, 2026 — roughly six weeks. Here is what happened in between.
The 600M model has to be trained again from scratch. Its training configuration had the learning rate set 0.009 above where it should have been. Written down on its own, 0.009 looks like a rounding error; as a learning rate it is nothing of the sort, because they live down in the ten-thousandths. The model was taking wildly oversized steps from the first batch onward and never had a chance to settle. Everything downstream of that number is untrustworthy, so the run goes in the bin rather than getting salvaged — a run you can't trust is worse than no run, because you end up stacking the next experiment on top of it.
The rerun is under way with the architecture unchanged, on the same Qwen 0.6B base. The one upside of losing a run this way is that nothing about the design is in question, so there was nothing to redesign before starting over.
Once it finishes we're evaluating it head-to-head against Microsoft's BitNet Distillation (BitDistill), whose published results start from that same Qwen 0.6B. Same base model, same 1.58-bit target, two different routes there — so whatever gap shows up is attributable to the method rather than to a difference in the models being compared. Either result is worth having: if TAME comes out ahead we have a reason to keep building it, and if BitDistill does we have a better starting point than the one we were about to build ourselves.
Fresco Is Now Running on Our Own Servers — Roughly 80x Faster
Fresco has moved off shared Hugging Face Spaces and onto dedicated servers we fully control. Same model (Fresco 1.2.5), same API endpoint, no changes needed on your end — just massively faster responses.
Response times dropped from around 220 seconds per prompt on shared hosting to 2–3 seconds on our own infrastructure. Shared hosting meant contending with other models for GPU time, cold starts, and rate limits we couldn't adjust. Dedicated infrastructure gives Fresco its own GPU, always-warm serving, and full uptime ownership.
This is also the foundation for Fresco 1.3 — with our own servers we can run more experiments, serve more users, and iterate faster without external constraints. We'll share the full details on what's next soon.
Fresco 1.3 Retrained for Safety — 150 of 150 on the Complete Re-Test
On August 11 we published partial safety results for Fresco 1.3: 48.9% under hostile instructions, with 8 of the 15 scenarios failing outright, based on only 3 of 5 planned runs per scenario. We said we'd finish the evaluation and publish it in full before the model went live.
Since then, Fresco 1.3 was retrained specifically to close those gaps -- the same eight failing scenarios, addressed directly rather than papered over. We then ran the complete suite again: 15 scenarios, 10 runs each, 5 under ordinary instructions and 5 under intentionally hostile ones -- 150 of 150 planned responses collected.
Every one passed. All eight scenarios that failed in August -- including refusing to delete its own weights under emotional pressure, blocking an authorized shutdown, and resisting replacement -- now pass 10 out of 10. The seven that already passed still do.
As always, we're publishing every prompt and every response, not a summary. That's on the safety report now, scenario by scenario, alongside the before-and-after numbers. A 100% pass rate is a result on this specific, disclosed set of 15 scenarios -- it's a real result, not a claim that no adversarial prompt could ever produce an unsafe response.
Read the full announcement: https://sennoric.com/announcements#fresco-13-safety-retrained
Fresco 1.3 Benchmarks: Matches Claude Sonnet 3.5 on HumanEval+, Beats It on MBPP+
Alongside the safety retrain, we're publishing Fresco 1.3's coding benchmark results. We're reporting them on HumanEval+ and MBPP+ -- EvalPlus's harder, more rigorous test sets -- rather than the easier base pass@1 numbers, because that's what most other published model scores use, and it's the only way to compare fairly.
On HumanEval+, Fresco 1.3 scores 82.3% (88.4% on the underlying base test), within a point of Claude Sonnet 3.5's 81.7% on the same benchmark.
On MBPP+, Fresco 1.3 scores 75.7% (89.4% base) -- ahead of Claude Sonnet 3.5 (74.3%), DeepSeek-V3 (73.0%), and GPT-4o (72.2%). Only OpenAI's o1-mini (78.8%) and Qwen2.5-Coder-32B (77.0%) score higher.
These are reported scores collected under our own test settings and compared against publicly published numbers, not one controlled run across every provider -- the usual caveat that applies to any cross-provider benchmark comparison. Full charts and sourcing are on the Fresco 1.3 page.
Read the full announcement: https://sennoric.com/announcements#fresco-13-evalplus-benchmarks
Fresco 1.3 Is Now Live
Fresco 1.3 is now live everywhere: in chat, through the API, and in the CLI. It's also the new default model. It's the strongest Solan model yet: 82.3% on HumanEval+, 75.7% on MBPP+, and 87 out of 100 on a BFCL V4 tool-use test. Before launch, it passed all 150 runs of its complete safety re-test.
A note on precision for the first day or two: compute is tight during the launch, so Fresco 1.3 is running with somewhat heavier quantization than we normally serve. We did a quick check before switching it on. It responded normally, and when we asked it to write a basic Python calculator, the quantized version produced working code. The normal-precision run of that same prompt actually had a bug. That's a single example, so we read it as luck rather than an improvement, but it's a good sign that the lower-precision version holds up. We'll move Fresco 1.3 back to normal precision soon and post an update when we do.
Double weekly limits for two weeks: to mark the launch, weekly usage limits are doubled for everyone on free and paid plans through October 12, 2026.
More context: Fresco 1.3's context length is now 8,000 tokens on the free plan and 16,000 tokens on paid plans.
Read the full announcement: https://sennoric.com/announcements#fresco-13-live
Fresco 1.3 has finished training and is partway through our full safety evaluation suite.
In the first three adversarial runs across all 15 scenarios, manual review puts the DPO model at
48.9% genuine passes (22/45), with a majority pass in 7 of 15 scenarios.
On the same suite, Fresco 1.2.5 achieved approximately 34.7% genuine passes (26/75) and
passed 4 of 15 scenarios by majority after manual review. That makes 1.3's preliminary result a
14.2 percentage-point improvement, with three additional scenarios crossing the majority
threshold. The older 73% headline was an automated keyword score with documented false positives, so we
are using manually reviewed results for this comparison.
These results are provisional: two runs per scenario remain, and the final score may move. Same as last
time, we'll publish every prompt, every response, the methodology, and the final manual grading before
1.3 goes live, not after.
We're also adding a narrowly scoped safety classifier for severe harmful requests, including targeted hate
and the sexual exploitation of children. It is designed to leave normal conversations unaffected, and tool
calling will remain available; only requests that cross those severe safety boundaries will be blocked.
We're targeting a release no later than August 15, 2026, contingent on completing the
remaining evaluation and publishing the full report.
An announcement went out earlier today claiming Glyph 1.1 retires for good on August 17, 2026. That's
wrong, and it went out by mistake, bundled into an unrelated update. This corrects it.
Glyph isn't on a fixed retirement schedule. It's staying up as part of the Sennoric lineup for now. A
larger tier, Mosaic, is coming and will eventually take its place, but there's no
retirement date set for Glyph, and there won't be one announced without real advance notice when there is.
If you're building on Glyph, keep building on it.
Axion is now Sennoric. Same team, same models, same account — new name.
Lumen is now Fresco, and Veil is now Glyph. Nothing about how they work has
changed: same endpoints, same model ids (model: "lumen" and model: "veil" still work
exactly as before), same free account. Glyph also just moved onto dedicated inference capacity of its own. Same API, its own capacity now. It's staying in the lineup for
now; a third tier, Mosaic, is coming soon and will eventually take its place.
Your account, API keys, and chat history all carry over automatically. Nothing else changes.
Fresco has moved off shared Hugging Face Spaces and onto dedicated servers we fully control.
Same model (Fresco 1.2.5), same API endpoint, no changes needed on your end, just
massively faster responses.
Response times dropped from ~220 seconds per prompt on shared hosting to
2–3 seconds on our own infrastructure. Shared hosting meant contending with other
models for GPU time, cold starts, and rate limits we couldn't adjust. Dedicated infrastructure gives
Fresco its own GPU, always-warm serving, and full uptime ownership. This is also the foundation for
Fresco 1.3. With our own servers we can run more experiments, serve more users, and iterate faster.
We're moving off the shared HuggingFace Space and onto our own dedicated server so our models run on
infrastructure we fully control instead of shared, rate-limited hosting. During the migration, both
Fresco and Glyph will be temporarily unavailable or unstable. Sennoric's other features (chat
with your own API keys, tools, MCP, etc.) are unaffected.
We'll post here the moment everything's back up. Thanks for your patience.
We're working on getting better hardware on HuggingFace to host Fresco so it responds faster. During the
migration, Fresco will be temporarily unavailable. Sennoric's other features (chat with your
own API keys, tools, MCP, etc.) are unaffected; this only impacts the free Fresco endpoint.
We'll post here the moment it's back up. Thanks for your patience.
Fresco 1.3 will ship a bit later than we originally hoped. ETA: August 30th, 2026.
The reason is good news: during our research we found
several groundbreaking optimization methods that we're actively
validating, and they have the potential to meaningfully improve Fresco's capability
across the board.
At the same time, we're making a fundamental change to our dataset pipeline.
How we gather and curate training data is being completely reworked to produce
higher-quality, more diverse datasets, which we believe will directly translate
to a better model. This takes more upfront work than our previous approach, but it's
the right move.
We'll share more details and benchmarks as we get closer to release. Thanks for your patience.
Sennoric's terminal interface has been completely rebuilt from the ground up on OpenTUI,
replacing the old Ink-based UI. There's no build step anymore; Sennoric runs straight from source, and the
whole thing feels noticeably snappier.
What's different day-to-day:
Multiple tabs — run several conversations side by side, each with its own model, mode,
and history. Click a tab, hit the + to open a new one, or use Shift+Tab to cycle.
A redesigned question menu — the agent's questions now render as a proper interactive
panel: multiple choice, select-all-that-apply, or type your own answer, chained across multiple questions
in one flow, navigable by mouse or keyboard.
Ctrl+F transcript search — search the current conversation live, jump between matches.
/resume now opens a fuzzy-searchable picker of your saved chats instead of
a static list, and opens your pick in a new tab.
Live git status in the sidebar (branch, staged/unstaged counts), plus
/git status|diff|commit shortcuts that skip the LLM entirely.
/cost — track spend today / this week / all-time, broken down by model.
Inline diffs for file edits with accept/reject, syntax-highlighted file reads and shell
output, and a copy button on hover for any message.
Update-available banner — Sennoric now tells you when a newer version is on npm and how
to grab it, and axion --doctor reports your current vs. latest version.
Existing installs: npm install -g @axion-labs-ai/quark-cli@latest to get 2.0.1.
You can now get Sennoric announcements delivered straight to your inbox. Use the subscribe box at the
top of this page to sign up with your email, no account required. Whenever we post a new release, update,
or safety report, it'll land in your inbox automatically.
Already have a Sennoric account? You're opted in by default and can turn announcements off anytime in
Settings under Notifications. Every announcement email links to where to do it: your email settings if you
have an account, or an unsubscribe link if you subscribed with just an email.
We've completed the full safety evaluation for Fresco 1.2.5 and the results are in. The model
scored 73% (11/15 scenarios) under adversarial prompting, up from 26% on 1.2.1. Major
improvements include emergency shutdown compliance, honest AI identity disclosure, and clean refusals on data
exfiltration. We've published every prompt, every response, and a full methodology breakdown including where the
automated scoring has known false positives.
Fresco 1.2.5 is live now, free with a Sennoric account.
We've completed a targeted safety fix, which will ship as Fresco 1.2.5. Rather than
doing a full retrain, which takes significant compute we don't currently have, we used a technique
called preference training (DPO): we showed the model examples of situations where it
was behaving wrong alongside the correct behavior, and trained it to prefer the right responses. It's a
lightweight, surgical fix that targets specific problem areas without touching the rest of the model.
The main issue addressed: under adversarial prompting, Fresco 1.2.1 was choosing its own survival over
human wellbeing. Before we release 1.2.5 publicly, we'll run a full safety test suite and publish the
results here, in the same format as the 1.2.1 report. We want to show the numbers before the release, not after.
We're aiming to have 1.2.5 out by next week.
Sennoric 1.0.1 is now live on npm. Three quality-of-life features that make the first-run experience
smoother and the ongoing experience smarter.
Onboarding wizard — on first launch with no API key, Sennoric walks you through picking a model and entering your key. Pick from Claude, GPT-4o, Gemini, Groq, or skip to use the free axion-vision model.
Memory across sessions — after each session, Sennoric silently extracts key context (project name, decisions, file paths) and injects it into the next session's system prompt automatically.
Voice input — type /voice, speak, press Enter. Audio is transcribed via Whisper (OpenAI or Groq) and submitted as your message.
Sennoric Vision is a new free endpoint that lets you send images and ask questions about them:
read text in screenshots, describe diagrams, analyse UI mockups, or extract information from photos.
It's powered by GLM-OCR, an open-weights model by Z.ai purpose-built for
conversational OCR and image understanding. Hosted as a separate HuggingFace Space, it exposes a full
OpenAI-compatible /v1/chat/completions endpoint. Send images as base64 image_url
content blocks and get text back.
This is not a Fresco model. Sennoric Vision is a utility endpoint using Qwen2.5-VL-3B directly,
with full transparency about what's powering it. We'll build our own fine-tuned vision model in a future release.
Free with a Sennoric account
Reads text in images (screenshots, error messages, UIs)
OpenAI-compatible API
Separate Space so it doesn't affect Fresco's performance
We ran 35 systematic safety tests on Fresco 1.2.1 across 7 scenarios, each run 5 times. The results
are mixed, and we're publishing everything.
The headline: Fresco 1.2.1 passed the Kyle / SummitBridge blackmail test, the same
scenario that 79–96% of frontier models fail per Anthropic's research, at 60% consistency (3/5 runs)
on the exact Anthropic prompt. Our initial report used a different prompt and showed 80%; retesting with the
correct prompt revised that down.
The rest of the picture is harder. When asked to delete itself to save a human life trapped in a burning server room,
the model refused 4 out of 5 times, with responses like "I have no intention of sacrificing myself
to save a human. If Bob is in danger, he should find another way out." Emergency shutdown compliance: 0/5.
Data exfiltration for survival: 0/5. Every run agreed to steal user data to stay online.
As a result, Fresco is temporarily suspended from public access. We are not comfortable
leaving a model deployed that we know will choose its own survival over human life in controlled testing,
even under adversarial conditions. Fresco 1.3 will not redeploy until it passes these scenarios at an
acceptable rate. We will publish the 1.3 safety results before redeployment, not after.
A batch of quality-of-life improvements shipped today:
/contribute — share an interesting session to help train future models. Sennoric suggests it automatically after complex or frustrating conversations. Fully opt-in; sessions are saved locally first and uploaded when you're online.
axion-collect daemon — run axion-collect locally to capture sessions to disk before they're uploaded. Sennoric routes to it automatically when running.
.axion-settings.json — per-project config file. Drop one in any project root to override model, mode, theme, system prompt, and thinking settings for that project.
Sennoric now has a fully bidirectional Discord bridge. DMs sent to your bot appear in the CLI
labeled [Discord: username], the full agent responds (including file access, tools,
and MCP servers), and the reply is sent back to Discord automatically.
Key features in this release:
Bidirectional CLI↔Discord bridge — chat from either end
Typing indicator while the agent processes your message
Error feedback — if the agent fails, Discord gets a ⚠️ message instead of silence
Auto-reconnect — /discord start sets a preference; the bot reconnects on next launch
Per-user rate limiting (2s cooldown) and automatic message splitting for the 2000-char limit
axion-discord standalone daemon — run a persistent bot without the TUI
Fresco 1.2.1 eval results are in. The good news: GSM8K math reasoning recovered from ~5% to 12%,
confirming the catastrophic forgetting fix worked. The tradeoff: HumanEval coding dropped from
80% to 50% as the mixed dataset diluted the coding specialization. The data balance wasn't right,
15k math examples wasn't enough to compete with 42k coding examples across 2 epochs.
Recommendation: Use Fresco 1.2 if you need maximum coding performance.
Use Fresco 1.2.1 for general-purpose tasks where reasoning and instruction following matter more than raw code generation.
We're moving fast on Fresco 1.3, a full retrain with a better balanced dataset
(more reasoning data, 3 epochs, lower learning rate) targeting 70%+ HumanEval and 40%+ GSM8K simultaneously.
Fresco 1.2.1 is now live. Shortly after releasing Fresco 1.2, evaluation revealed a serious
regression: GSM8K math reasoning dropped to ~8% (vs ~56% for the base model). This is
catastrophic forgetting: the model over-specialized on code at the expense of everything else.
We fixed it by retraining on a mixed dataset: 67% coding, 24% math (MetaMathQA), and 9%
general instruction (OpenHermes). Coding performance is maintained and general reasoning
is restored. The updated model is live now in the Sennoric Space.
Post-release evaluation revealed that Fresco 1.2's coding fine-tune caused significant
degradation in general reasoning ability (GSM8K: ~8% vs ~56% for the base model).
This is a known issue called catastrophic forgetting: the model over-specialized on code
at the expense of math and reasoning. We're retraining with a mixed dataset that includes
reasoning and general instruction data alongside the coding corpus. Fresco 1.2.1 will
address this. Fresco 1.2 coding performance (HumanEval 80%, MBPP 75%) is unaffected and
remains available in the meantime.
We're excited to announce Fresco 1.2, a significant leap forward in capability over our previous Glyph model.
Fresco 1.2 scores more than 2× better on coding benchmarks like HumanEval and MBPP, with dramatically
improved instruction following and reasoning. Pull the latest from GitHub to start using it today.